Models
VideoPoet is a Google Research video model that was not released publicly. It isn't available on Muvi; this page explains it and points to models you can use.
This guide is available in English.
Published by Google Research in late 2023, VideoPoet was designed to demonstrate that a single autoregressive language model could perform diverse video generation and editing tasks without specialized per-task heads:
1. Decoder-Only Autoregressive Architecture: VideoPoet treats video and audio as discrete tokens. Using an architecture similar to LLMs like PaLM, it generates video frames token-by-token from left-to-right and past-to-future. 2. MAGVIT-V2 Video Tokenization: VideoPoet relied on MAGVIT-V2, an advanced neural video tokenizer that maps variable-resolution video clips into a compact vocabulary of discrete spatial-temporal tokens. 3. Multimodal Task Unification: Because it operates as an LLM over tokens, VideoPoet natively handles text-to-video, image-to-video, video editing, video-to-audio sound generation, and interactive camera direction. 4. The Transition to Modern Production Diffusion: While autoregressive video generation demonstrated strong conceptual versatility, modern high-resolution diffusion transformers (like Google Veo 3.1 and ByteDance Seedance 2.0) have emerged as the industry standard for production rendering. Diffusion models deliver superior fine texture resolution and eliminate token quantization blur, and Google Veo 3.1 scales to 4K Ultra HD.
Evaluate these sample clips, which demonstrate cinematic camera trajectories, structural geometric stability, and dynamic lighting:
“Low-angle cinematic tracking shot through a cyberpunk alleyway, volumetric purple and amber light casting realistic reflections across wet pavement puddles, subtle atmospheric steam rising from storm drains, 4K resolution, 24fps”
Notice the clean specular highlights and authentic water puddle reflections. Where tokenized autoregressive models often produce subtle grid artifacts and quantization noise along high-frequency edges, modern studio diffusion preserves razor-sharp clarity.
“High-speed orbital tracking shot around a futuristic hover vehicle soaring through a misty alpine valley at sunrise, crisp volumetric sun rays filtering through pine canopies, photorealistic kinetic motion blur, 4K resolution”
This clip stresses high-speed spatial parallax and volumetric sunlight scattering. The hover vehicle maintains mathematically solid geometry without the frame-to-frame drifting common in early autoregressive research experiments.
Moving from academic research papers to Muvi Studio empowers digital storytellers, marketing agencies, and filmmakers:
1. Immediate Browser Access to Google Veo 3.1: Instead of waiting for Google Research demos to become public tools, Muvi provides immediate browser access to Google DeepMind's flagship Veo 3.1 model. 2. Native Synchronized Audio on Google Veo 3.1: Just as VideoPoet explored video-to-audio synthesis, Google Veo 3.1 brings true synchronized dialogue, sound effects, and atmospheric music directly into production video files. 3. Up to 15-Second Continuous Takes with Seedance 2.0: Need extended single-shot takes? ByteDance Seedance 2.0 on Muvi delivers continuous takes of 4 to 15 seconds with persistent character identity and natural motion pacing. 4. In-Frame Typography with Alibaba Wan 3.0 Prime: Synthesize readable brand logos, packaging labels, and signage directly in your video without post-production graphic overlays. 5. Free Starter Coins & Clear Coin Costs: Test prompts with free starter coins upon signup. Every model shows its coin cost before you generate, and plans and coin packs are listed on the pricing page.
| Feature / Metric | Google VideoPoet (Research) | Google Veo 3.1 (Live on Muvi) | ByteDance Seedance 2.0 | Alibaba Wan 3.0 Prime |
|---|---|---|---|---|
| Core Architecture | Autoregressive LLM + MAGVIT-V2 | DeepMind Spatio-Temporal Diffusion | Flow-Matching Multimodal DiT | Spatio-Temporal DiT |
| Max Native Duration | ~5s (autoregressively extendable) | 4s, 6s, 8s native single-shot | 4 to 15 seconds continuous | 2s – 30s continuous |
| Output Resolutions | ~512p to 720p native | 720p, 1080p Full HD, 4K Ultra HD | 480p, 720p, 1080p | 480p, 720p, 1080p Full HD |
| Native Audio Generation | Separate SoundStream tokens | Native synchronized dialogue & foley | Audio supported | Optional sound |
| Availability | Academic research paper | Live on Muvi | Live on Muvi | Live on Muvi |
| Commercial Rights | None (research paper) | Full Commercial Rights on Pro and higher plans | Full Commercial Rights on Pro and higher plans | Full Commercial Rights on Pro and higher plans |
How the models compare on visual fidelity, motion dynamics, and production readiness:
| Dimension | Google VideoPoet | Google Veo 3.1 | ByteDance Seedance 2.0 |
|---|---|---|---|
| Audio-Video Integration | Dual-token autoregression | Native synced audio | Audio supported |
| Generation Latency | Sequential token decoding (slow) | Cloud diffusion (typically minutes) | Cloud diffusion (typically minutes) |
| Accessibility | Not publicly available | Free starter coins / paid plans | Free starter coins / Pro plan (2.0 Fast on all plans) |
Use these studio-crafted prompt templates in Muvi Studio to achieve rich narrative depth and cinematic realism:
Cinematic medium close-up of a lone scientist stationed inside an Arctic geodesic observatory during a polar blizzard. Green aurora borealis glows through the frost-rimmed glass dome. The scientist speaks calmly into a field recorder: "Ice core sample twelve shows anomalous atmospheric density." Native audio: resonant voice acoustics, howling blizzard winds rattling against the glass dome, electronic equipment hum. 1080p, 24fps.Dynamic tracking shot following a competitive breakdancer executing an explosive windmill into a hand-spin on a sunlit street court. Realistic fabric drag on the cotton hoodie, fine dust particles kicking up from the asphalt, golden sunlight catching sweat droplets. Continuous 12-second single take, flawless limb anatomy throughout the rotations, photorealistic 1080p, 30fps.Studio commercial macro shot of a sleek matte black bottle of craft tonic water on a slab of rough slate. Crisp, perfectly legible embossed silver typography on the label reads "VIDEOPoet BOTANICALS". Fine cold vapor drifts from the cap, condensation droplets slowly rolling down the glass with authentic refractive physics, controlled studio softbox lighting, 1080p resolution.Google VideoPoet is a multimodal research model developed by Google Research that uses an autoregressive language model architecture to generate video and audio tokens. It was published as an academic paper in December 2023 and was never launched as an independent public product. VideoPoet is not available on Muvi. Instead, Muvi provides a zero-setup cloud studio with production models like Google Veo 3.1, ByteDance Seedance 2.0, and Alibaba Wan 3.0 Prime, and new users get free starter coins to try lower-cost models such as Alibaba Wan 3.0 Prime and Seedance 2.0 Fast.
Most contemporary AI video models (like Sora or Stable Video Diffusion) use diffusion processes that progressively denoise latents. VideoPoet instead treated video generation like an LLM (such as GPT or Gemini), using discrete MAGVIT-V2 spatial-temporal video tokens that the model predicts sequentially from past to future.
While VideoPoet demonstrated that language models could generate synchronized audio and video tokens, Google Veo 3.1 brings this capability to full commercial production. Veo 3.1 synthesizes photorealistic 1080p and 4K video with native synchronized dialogue, natural character lip-syncing, foley effects, and environmental room acoustics.
Yes. Pro, Ultra and Business plans on Muvi include 100% watermark-free exports with full commercial usage rights. Resolution depends on the model: Veo 3.1 exports 720p, 1080p or 4K, while Seedance 2.0 and Wan 3.0 Prime export up to 1080p.
One coin balance works across the video models on Muvi, including Google Veo 3.1, ByteDance Seedance 2.0 and Alibaba Wan 3.0 Prime, and each model shows its coin cost before you generate. Some models need a specific plan: Seedance 2.0 requires Pro and Gemini Omni Flash requires Ultra. Current plans and coin packs are listed on the pricing page.
Discover the full suite of foundation video models:
VideoPoet is a Google Research video model that was not released publicly. It isn't available on Muvi; this page explains it and points to models you can use.